Donato Capitella
334 خطوط
12 KiB
Python
334 خطوط
12 KiB
Python
#!/usr/bin/env python3
|
|
import subprocess, time, json, sys, os, requests, argparse
|
|
from pathlib import Path
|
|
|
|
try:
|
|
import bench_utils
|
|
except ImportError:
|
|
sys.path.append(str(Path(__file__).parent))
|
|
import bench_utils
|
|
|
|
|
|
# =========================
|
|
# ⚙️ GLOBAL SETTINGS
|
|
# =========================
|
|
|
|
try:
|
|
import models
|
|
except ImportError:
|
|
# If running locally and models.py is in ../scripts?
|
|
# Or if running in /opt where models.py is alongside.
|
|
# We will try adding current dir to path just in case
|
|
sys.path.append(os.getcwd())
|
|
try:
|
|
import models
|
|
except ImportError:
|
|
# Fallback for local structure: assuming this is in benchmarks/ and models is in scripts/
|
|
sys.path.append(str(Path(__file__).parent.parent / "scripts"))
|
|
import models
|
|
|
|
# Import from shared config
|
|
MODEL_TABLE = models.MODEL_TABLE
|
|
MODELS_TO_RUN = models.MODELS_TO_RUN
|
|
GPU_UTIL = models.GPU_UTIL
|
|
OFF_NUM_PROMPTS = models.OFF_NUM_PROMPTS
|
|
OFF_FORCED_OUTPUT = models.OFF_FORCED_OUTPUT
|
|
DEFAULT_BATCH_TOKENS = models.DEFAULT_BATCH_TOKENS
|
|
|
|
# Fallbacks
|
|
FALLBACK_INPUT_LEN = 1024
|
|
FALLBACK_OUTPUT_LEN = 512
|
|
|
|
RESULTS_DIR = Path("benchmark_results")
|
|
RESULTS_DIR.mkdir(exist_ok=True)
|
|
|
|
|
|
# =========================
|
|
# UTILS
|
|
# =========================
|
|
|
|
def log(msg): print(f"\n[BENCH] {msg}")
|
|
|
|
def get_gpu_count():
|
|
try:
|
|
# Using rocm-smi --showid to list GPUs.
|
|
# Output format: "GPU[0] : Device Name: ..."
|
|
res = subprocess.run(["rocm-smi", "--showid"], stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True)
|
|
if res.returncode == 0:
|
|
return 1 # Force return 1 for Strix Halo APU
|
|
else:
|
|
log("rocm-smi failed, defaulting to 1 GPU (Hardcoded Fallback)")
|
|
return 1
|
|
except Exception as e:
|
|
log(f"Error detecting GPUs: {e}, defaulting to 1 GPU")
|
|
return 1
|
|
|
|
def kill_vllm():
|
|
subprocess.run("pgrep -f 'vllm serve' | xargs -r kill -9",
|
|
shell=True, stdout=subprocess.DEVNULL, stderr=subprocess.DEVNULL)
|
|
time.sleep(5)
|
|
|
|
def nuke_vllm_cache():
|
|
cache = Path.home() / ".cache" / "vllm"
|
|
if cache.exists():
|
|
try:
|
|
subprocess.run(["rm", "-rf", str(cache)], check=True)
|
|
cache.mkdir(parents=True, exist_ok=True)
|
|
time.sleep(2)
|
|
except: pass
|
|
|
|
def get_dataset():
|
|
data_path = Path("ShareGPT_V3_unfiltered_cleaned_split.json")
|
|
if data_path.exists(): return str(data_path)
|
|
|
|
log("Downloading ShareGPT dataset...")
|
|
url = "https://huggingface.co/datasets/anon8231489123/ShareGPT_Vicuna_unfiltered/resolve/main/ShareGPT_V3_unfiltered_cleaned_split.json"
|
|
try:
|
|
r = requests.get(url, stream=True, timeout=15)
|
|
r.raise_for_status()
|
|
with open(data_path, 'wb') as f:
|
|
for chunk in r.iter_content(chunk_size=8192): f.write(chunk)
|
|
return str(data_path)
|
|
except Exception as e:
|
|
log(f"WARNING: ShareGPT download failed ({e}). using RANDOM.")
|
|
return None
|
|
|
|
|
|
|
|
def get_model_args(model, tp_size, overrides=None):
|
|
config = MODEL_TABLE.get(model, {"max_num_seqs": "32"})
|
|
overrides = overrides or {}
|
|
|
|
# Allow per-model GPU utilization override
|
|
util = overrides.get("gpu_util", config.get("gpu_util", GPU_UTIL))
|
|
max_seq_override = overrides.get("max_num_seqs", config.get("max_num_seqs", "32"))
|
|
|
|
cmd = [
|
|
"--model", model,
|
|
"--gpu-memory-utilization", str(util),
|
|
"--dtype", "auto",
|
|
"--tensor-parallel-size", str(tp_size),
|
|
"--max-num-seqs", str(max_seq_override)
|
|
]
|
|
|
|
# Optional: if a model really needs a hard limit, we can still support "ctx" in config,
|
|
# but by default we rely on auto.
|
|
if "ctx" in overrides or "ctx" in config:
|
|
cmd.extend(["--max-model-len", str(overrides.get("ctx", config.get("ctx")))])
|
|
|
|
if config.get("trust_remote"): cmd.append("--trust-remote-code")
|
|
if config.get("enforce_eager"): cmd.append("--enforce-eager")
|
|
|
|
return cmd
|
|
|
|
def run_throughput(model, tp_size, backend_name="Default", output_dir=RESULTS_DIR, extra_env=None, overrides=None):
|
|
if tp_size not in MODEL_TABLE[model]["valid_tp"]: return
|
|
overrides = overrides or {}
|
|
|
|
model_safe = model.replace("/", "_")
|
|
output_dir_path = Path(output_dir)
|
|
output_dir_path.mkdir(parents=True, exist_ok=True)
|
|
|
|
tag = overrides.get("tag", "").strip()
|
|
tag_suffix = f"_{tag}" if tag else ""
|
|
output_file = output_dir_path / f"{model_safe}_tp{tp_size}{tag_suffix}_throughput.json"
|
|
|
|
if output_file.exists():
|
|
log(f"SKIP {model} (TP={tp_size} | {backend_name})")
|
|
return
|
|
|
|
dataset_path = get_dataset()
|
|
dataset_args = ["--dataset-name", "sharegpt", "--dataset-path", dataset_path] if dataset_path else ["--input-len", "1024"]
|
|
|
|
# Retrieve Model-Specific Batch Tokens
|
|
batch_tokens = str(overrides.get("max_tokens", MODEL_TABLE[model].get("max_tokens", DEFAULT_BATCH_TOKENS)))
|
|
|
|
log(f"START {model} (TP={tp_size} | {backend_name}) [Batch: {batch_tokens}]...")
|
|
kill_vllm()
|
|
nuke_vllm_cache()
|
|
|
|
cmd = ["vllm", "bench", "throughput"] + get_model_args(model, tp_size, overrides)
|
|
cmd.extend([
|
|
"--num-prompts", str(OFF_NUM_PROMPTS),
|
|
"--max-num-batched-tokens", batch_tokens,
|
|
"--output-len", OFF_FORCED_OUTPUT,
|
|
"--output-json", str(output_file),
|
|
"--disable-log-stats"
|
|
])
|
|
cmd.extend(dataset_args)
|
|
|
|
# Force Attention Backend via CLI if ROCm-Attn
|
|
if backend_name == "ROCm-Attn":
|
|
cmd.extend(["--attention-backend", "ROCM_ATTN"])
|
|
|
|
# ENV Setup: Global + Model Specific
|
|
env = os.environ.copy()
|
|
env["VLLM_DISABLE_COMPILE_CACHE"] = "1"
|
|
|
|
# Inject model specific env vars (e.g. for AWQ)
|
|
model_env = MODEL_TABLE[model].get("env", {})
|
|
env.update(model_env)
|
|
|
|
# Extra Env
|
|
if extra_env:
|
|
env.update(extra_env)
|
|
|
|
try:
|
|
subprocess.run(cmd, check=True, env=env)
|
|
except:
|
|
log(f"ERROR: Failed {model} [{backend_name}]")
|
|
|
|
|
|
def print_summary(tps):
|
|
print(f"\n{'MODEL':<40} | {'TP':<2} | {'Tag':<15} | {'Triton':<8} | {'ROCm':<8}")
|
|
print("-" * 92)
|
|
|
|
for m in MODELS_TO_RUN:
|
|
msafe = m.replace("/", "_")
|
|
name_cell = m.split('/')[-1]
|
|
|
|
for tp in tps:
|
|
if tp not in MODEL_TABLE[m]["valid_tp"]: continue
|
|
|
|
prefix = f"{msafe}_tp{tp}"
|
|
|
|
tags = set()
|
|
for p in RESULTS_DIR.glob(f"{prefix}*_throughput.json"):
|
|
name_part = p.name[len(prefix):-len("_throughput.json")]
|
|
tag = name_part.lstrip("_")
|
|
tags.add(tag)
|
|
|
|
for p in Path("benchmark_results_rocm").glob(f"{prefix}*_throughput.json"):
|
|
name_part = p.name[len(prefix):-len("_throughput.json")]
|
|
tag = name_part.lstrip("_")
|
|
tags.add(tag)
|
|
|
|
if not tags:
|
|
tags.add("") # Default empty tag if no files found
|
|
|
|
for tag in sorted(list(tags)):
|
|
tag_suffix = f"_{tag}" if tag else ""
|
|
|
|
# Default
|
|
try:
|
|
p1 = RESULTS_DIR / f"{prefix}{tag_suffix}_throughput.json"
|
|
if p1.exists():
|
|
d1 = json.loads(p1.read_text())
|
|
val1 = f"{d1.get('tokens_per_second', 0):.1f}"
|
|
else:
|
|
val1 = "N/A"
|
|
except: val1 = "N/A"
|
|
|
|
# ROCm
|
|
try:
|
|
p2 = Path("benchmark_results_rocm") / f"{prefix}{tag_suffix}_throughput.json"
|
|
if p2.exists():
|
|
d2 = json.loads(p2.read_text())
|
|
val2 = f"{d2.get('tokens_per_second', 0):.1f}"
|
|
else:
|
|
val2 = "N/A"
|
|
except: val2 = "N/A"
|
|
|
|
display_tag = tag if tag else "(Default)"
|
|
print(f"{name_cell:<40} | {tp:<2} | {display_tag:<15} | {val1:<8} | {val2:<8}")
|
|
|
|
print("-" * 92)
|
|
|
|
if __name__ == "__main__":
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument("--tp", type=int, nargs="+", default=[1])
|
|
parser.add_argument("--tui", action="store_true", help="Launch interactive configuration UI")
|
|
args = parser.parse_args()
|
|
|
|
gpu_count = get_gpu_count()
|
|
log(f"Detected {gpu_count} AMD GPU(s)")
|
|
|
|
valid_tp_args = [t for t in args.tp if t <= gpu_count]
|
|
if not valid_tp_args:
|
|
log(f"Requested TP={args.tp} but only {gpu_count} GPU(s) detected. Nothing to run.")
|
|
sys.exit(0)
|
|
|
|
selected_models = MODELS_TO_RUN
|
|
|
|
if args.tui:
|
|
# TUI Model Selection
|
|
checklist_args = [
|
|
"--clear", "--backtitle", "AMD vLLM Benchmark Launcher",
|
|
"--title", "Model Selection",
|
|
"--checklist", "Select models to benchmark:", "20", "65", "10"
|
|
]
|
|
|
|
for m in MODELS_TO_RUN:
|
|
m_name = m.split("/")[-1]
|
|
# All selected "on" by default
|
|
checklist_args.extend([m, m_name, "on"])
|
|
|
|
choice = bench_utils.run_dialog(checklist_args)
|
|
|
|
if choice is None:
|
|
subprocess.run(["clear"])
|
|
print("Cancelled by user.")
|
|
sys.exit(0)
|
|
|
|
# Parse space-separated quoted output from dialog checklist
|
|
import shlex
|
|
selected_models = [m for m in shlex.split(choice)]
|
|
|
|
if not selected_models:
|
|
subprocess.run(["clear"])
|
|
print("No models selected. Exiting.")
|
|
sys.exit(0)
|
|
|
|
kill_vllm()
|
|
for tp in valid_tp_args:
|
|
for m in selected_models:
|
|
overrides = {}
|
|
if args.tui:
|
|
config = MODEL_TABLE.get(m, {})
|
|
default_seqs = config.get("max_num_seqs", "32")
|
|
default_tokens = config.get("max_tokens", DEFAULT_BATCH_TOKENS)
|
|
default_util = config.get("gpu_util", GPU_UTIL)
|
|
default_ctx = config.get("ctx", "auto")
|
|
|
|
form_args = [
|
|
"--clear", "--backtitle", f"AMD vLLM Benchmark Configuration (TP: {tp})",
|
|
"--title", f"Tune Parameters: {m.split('/')[-1]}",
|
|
"--form", "Edit the options below. Leave tag empty for no suffix.",
|
|
"15", "70", "5",
|
|
"Max Concurrent Seqs:", "1", "1", str(default_seqs), "1", "25", "15", "0",
|
|
"Max Batched Tokens:", "2", "1", str(default_tokens), "2", "25", "15", "0",
|
|
"GPU Utilization (0-1):", "3", "1", str(default_util), "3", "25", "15", "0",
|
|
"Max Context Length:", "4", "1", str(default_ctx), "4", "25", "15", "0",
|
|
"Filename Tag (Optional):", "5", "1", "", "5", "25", "15", "0"
|
|
]
|
|
|
|
form_res = bench_utils.run_dialog(form_args)
|
|
if form_res is None:
|
|
subprocess.run(["clear"])
|
|
print(f"Skipping {m} (TP={tp}) due to user cancellation.")
|
|
continue
|
|
|
|
lines = form_res.splitlines()
|
|
if len(lines) >= 5:
|
|
overrides["max_num_seqs"] = lines[0].strip()
|
|
overrides["max_tokens"] = lines[1].strip()
|
|
overrides["gpu_util"] = lines[2].strip()
|
|
|
|
ctx_val = lines[3].strip()
|
|
if ctx_val and ctx_val.lower() != "auto":
|
|
overrides["ctx"] = ctx_val
|
|
|
|
overrides["tag"] = lines[4].strip()
|
|
|
|
# 1. Default (Triton)
|
|
run_throughput(m, tp, "Default", RESULTS_DIR, overrides=overrides)
|
|
|
|
# 2. ROCm Attention
|
|
# We force this via CLI argument --attention-backend ROCM_ATTN below
|
|
# No specific env vars needed if forcing backend.
|
|
rocm_env = {}
|
|
print(f"[DEBUG] Forcing ROCm Env: {rocm_env} + CLI: --attention-backend ROCM_ATTN")
|
|
run_throughput(m, tp, "ROCm-Attn", "benchmark_results_rocm", rocm_env, overrides=overrides)
|
|
|
|
print_summary(valid_tp_args)
|